The Research Assistant is a conversational AI agent that has access to all of a project's interview data — transcripts, analysis pages, comparisons, segments, and statistics — and answers questions about it in natural language. Replies come with cited respondents, structured artifacts (quote collections, data tables, theme summaries, charts, markdown reports), and full conversation history.
The Research Assistant replaces the previous AI Search feature. The same tab in the dashboard now hosts the new experience; the URL bookmark for AI Search continues to work.
Beta — the Research Assistant is currently in Beta. It is fully usable today, and we are actively refining it, so some details may change from release to release. A small Beta label sits next to the tab and inside the chat header to remind you.
What it can do
- Multi-turn conversation — ask follow-up questions, dig into a finding, switch topics, all within the same thread. The assistant remembers what you've already discussed and which respondents you've cited.
- Iterative exploration — start with an overview, then ask for more depth. Fast first answers come from the analysis Tellet has already prepared, then it searches the full transcripts when you ask to go deeper ("find me more", "show me specific quotes", "who disagreed?").
- Knows where to look — the assistant chooses which part of your data to consult: overall and per-question analysis, comparisons, categories, full transcripts, individual interviews, or live statistics. Just ask the question; you don't need to know where the answer lives.
- Cited answers — every claim is backed by a respondent reference like
[R7] that opens the source transcript in a side panel. Hover or click to read the full context.
- Structured artifacts — when your question warrants more than prose, the assistant generates a side-panel artifact you can read, copy, or export:
- Quote collection with audio/video media when available
- Data table with rows of statistics
- Theme summary with named themes and supporting evidence
- Bar chart for distributions
- Markdown report for executive summaries
- Filter scope — each thread is scoped to a filter (an audience segment) or to "all respondents". Switching filters creates a new thread so the assistant doesn't accidentally mix segments mid-conversation.
- Thread management — every chat is saved as a thread you can return to later, share read-only with teammates (via the share toggle), or delete.
- Background processing — switching to another tab (Questions, Transcripts, etc.) while the assistant is working does NOT interrupt the response. When you return, the completed response will be waiting for you.
- Progress visibility — a global progress bar at the top of the page shows when the Research Assistant (or analysis/comparison) is running, even if you navigate to a different tab.
- Export — copy any artifact to the clipboard, download a chart artifact as a PNG image, or download the entire session as a Markdown document for reports or client presentations.
- 64+ languages — the assistant detects the language of your question and responds in that same language. When a respondent answered in a different language, their quote is shown in your project's analysis language, with their original words preserved underneath so you can always check the verbatim source.
Accessing the Research Assistant
Navigate to a project, then click the Research Assistant tab in the secondary navigation (it sits alongside Overview, Questions, Transcripts, and Compare). The tab appears once your project is marked Completed (from the project status menu) — while responses are still coming in, the analysis keeps changing, so the assistant stays hidden until you close data collection. It is available on laptop and desktop screens (it is hidden on small mobile screens).
How the assistant finds answers
The assistant balances speed with depth. It starts with the fast path — the analysis Tellet has already prepared (curated themes, quotes, statistics, and patterns) for broad questions, or one question's analysis for a question-specific ask. When the prepared analysis isn't enough — because you asked for more, want to go deeper, or the topic isn't covered — it searches the full text of every interview. That leaves nothing out, but takes a little longer.
You can trigger the deeper search at any time — just say "search the transcripts" or "find me more quotes."
Asking a question
The empty state shows category cards with suggested questions tailored to your project:
- Explore themes — questions about main findings, patterns, themes
- Find quotes — questions that surface specific verbatim quotes
- Compare segments — cross-segment questions when you have filters or saved comparisons
- Generate reports — executive summary, narrative, or full report questions
Click any suggestion to send it as your first message, or type your own question into the input at the bottom.
What kinds of questions work well
First-pass questions (fast):
- "What are the main themes from this study?"
- "Give me an executive summary"
- "What did respondents say about question 3?"
- "How many women mentioned quality?"
- "Compare the male and female segments"
Follow-up exploration (deeper search):
- "Find me more surprising responses beyond those"
- "Show me specific quotes that support theme 1"
- "Who disagreed with the main findings?"
- "Dig deeper into what respondents said about pricing"
- "Search the transcripts for mentions of sustainability"
Respondent-level questions:
- "Show me respondent 5's full interview"
- "What did respondent 7 say about onboarding?"
Verification and challenge:
- "Prove it — show me what all respondents said about this"
- "Are there any views that contradict that theme?"
- "How many respondents actually mentioned this?"
Reports and exports:
- "Generate an executive summary with key quotes for a client presentation"
- "Create a detailed report on this study"
How questions are handled
You don't need to tell the assistant where to look — it reads your question and chooses the right source:
- "What are the themes / key findings / summary?": the overall analysis
- "What did respondents say about question X?": that question's analysis
- "Find me more / dig deeper / show me quotes": a search of the full transcripts
- "Who disagreed? / Any contradictions? / Did anyone mention X?": a search of the full transcripts
- "Show me respondent X's full interview": that respondent's complete interview
- "How many people… / What % of…": live statistics
- "Compare segment A and segment B": an existing comparison
- "What categories were assigned to question Q?": that question's categories
It figures this out from the question and explains its reasoning when that's helpful.
What the assistant can draw on
- Your interview design — goal, audience, questions, screening criteria, analysis language, and metadata fields.
- The overall analysis — executive summary, themes, cross-question patterns, recommendations, representative quotes, sample profile, notable visuals, per-question summaries, and statistics.
- Per-question analysis — themes, quotes, opportunities, and visuals for one question.
- The full transcripts — the complete text of every interview, searchable by meaning, for deeper exploration, specific quotes, counter-evidence, and uncovered topics.
- Individual interviews — one respondent's complete set of answers, in order.
- Your saved segments and comparisons — audience segments by name, and any comparison already run on the Compare tab.
- Categories — the category counts and definitions for a question.
- Live statistics — counts, averages, cross-tabulations, group comparisons, and distributions, calculated on demand.
Every artifact is checked against the underlying data before it's shown to you.
Reading the response
While the assistant is thinking, the chat shows a "Thinking..." indicator followed by a progress stepper that shows each step as it runs and completes. When it finishes, a short summary shows how many steps were used and the elapsed time (e.g., "3 tools used (17s)").
Citations — markers like [R7] in the response link to the respondent's transcript. Click one to open that respondent's transcript in a side panel and see the full context for their answers.
Artifacts — when the assistant generates a structured output (3+ quotes, a table, a chart, a report), the artifact panel slides in from the right. When there are multiple artifacts, a dropdown selector at the top lets you switch between them. The chat itself stays clean with a short intro pointing you to the artifact.
Source question — when an artifact is based on a specific interview question, the artifact header shows which question the data came from (e.g., "Based on Q3: 'Why did you choose that service?'").
Analytical rigour — the assistant notes sample sizes, hedges findings proportional to evidence strength ("3 of 10 respondents" vs. "respondents said"), and flags when its answer is based on a sample of the data rather than every interview.
Answer integrity — the assistant only reports what is actually in your data. It will not invent respondents or made-up personas, and it will not present a chart or figure it did not genuinely produce from your interviews. Every quote and citation traces back to a real respondent you can open and read.
Truncation — if a response runs long and is cut short, an indicator appears with a "Continue" button that asks the assistant to pick up where it left off.
Thread management
Threads appear in the left sidebar. Each thread shows its title (auto-generated from the first message in your language), the relative date of the last activity, and the message count.
- New thread — click the "New Thread" button at the top of the sidebar
- Switch threads — click a thread row to load its history
- Delete — hover over a thread, click the trash icon, confirm in the dialog
- Share — open a thread, click the share toggle in the header (read-only access for anyone with the URL)
Threads are scoped to your active project filter. If you switch the dashboard filter while in the Research Assistant, you'll see a different set of threads — the ones created under that filter.
Exporting results
Single artifact
Click the Copy button in the artifact panel header to copy the active artifact as formatted text suitable for pasting into Google Docs, email, or PowerPoint.
When the artifact is a chart, a Download button next to Copy saves it as a high-resolution PNG image — ready to drop straight into a slide deck or report.
Entire session
Click the Download button (next to Copy) to download the whole session — every artifact in order — as a single Markdown document. The filename includes the export date.
Performance
Response times vary by project size and network conditions. As a rough guide to relative speed:
- Overview / themes / summary: Fastest — answered from prepared analysis
- Question-specific analysis: Fast — answered from that question's analysis
- Full respondent interview: Fast — loads one interview
- Transcript search (follow-up): Slower — searches the full text of every interview
- Executive summary / report: Slowest — assembles a longer, structured output
Follow-up questions within the same session are usually faster than the first one.
Limitations and known gaps
- Languages — the assistant supports questions and responses in 64+ languages but the dashboard interface (button labels, tooltips, error messages) is currently English and Dutch only.
- Rate limit — to keep the assistant responsive for everyone, you can send up to 30 messages per hour.
- Long threads — in a very long conversation the assistant focuses on your most recent messages; earlier ones stay saved and visible but may not inform later answers.
- Citation accuracy — the assistant catches fabricated quotes and unknown respondent references, but for very short responses or paraphrased citations there may be edge cases. Always click through to the transcript to confirm any claim you'll show to a client.
- Deeper search by meaning — the full-transcript search matches on meaning, so it can miss an exact phrase that's worded very differently from your query. For precise phrase matching, rephrase with more specific terms.
- Counting by topic is approximate — "how many respondents mentioned X?" is estimated from the full-transcript search (the number of distinct respondents it surfaces). For exact counts on metadata like age or gender, ask for a statistic directly.
Troubleshooting
"Something went wrong" — the assistant hit a temporary problem (it retries automatically behind the scenes) or you've reached the hourly message limit. Click Dismiss and try again.
"No quotes found for this query" — the assistant genuinely found nothing for that request. Try rephrasing the question or relaxing the scope (e.g., remove the filter).
Empty artifact panel — if you opened the panel but see nothing, the active artifact may have been deleted by a follow-up message. Use the dropdown selector to switch to a different artifact.
Slow responses — searching the full transcripts is the slowest step. The assistant uses it only when you ask for deeper exploration or when the prepared analysis isn't enough. While a response is still being generated, the progress bar at the top of the page shows it's still working.
Response cut off — if a response ends mid-sentence, click the "Continue" button that appears below the message to ask the assistant to finish.